[SCIP] SCIP speed-up for .LP files
Ambros Gleixner
gleixner at zib.de
Fri Jun 15 10:37:59 CEST 2018
Hi Pierre and Marcus,
The attached problem was small and also SCIP/SoPlex solves it easily.
Marcus sent me a larger instance. These seem pure feasibility problems
with a tight dual bound of 0, but hard combinatorial structure.
Marcus: Maybe adding an artificial objective function helps to direct
the search, certainly turning off separation and maybe also
strongbranching should speed things up. But potentially SCIP is missing
a heuristic that makes the commercial solvers more successful.
Best,
Ambros
Am 15.06.2018 um 02:54 schrieb Pierre Le Bodic:
> Hi Marcus,
>
> I gave a try to your problem. Using scip 5.0.1 and gurobi 8.0.0 as the
> LP solver, I get a feasible solution in less than a second. Log
> attached. Please let me know if you can't replicate this.
>
> Kind regards,
> Pierre
>
> On 15 June 2018 at 08:38, Ambros Gleixner <gleixner at zib.de
> <mailto:gleixner at zib.de>> wrote:
>
> Hi Marcus,
>
> 28456 variables are in general not particularly large, but for hard
> problems, also medium-sized instances can become very difficult.
>
> Gurobi is known to be stronger than SCIP, but it could be that with
> the right parameters also SCIP can solve the instances. As a
> starter try to change emphasis settings, one or more of
>
> set presolving emphasis aggressive
> set separating emphasis {off,fast,aggressive}
> set heuristics emphasis aggressive
> set emphasis {feasibility,optimality,hardlp}
>
> and/or look at the statistics via "display statistics" to find
> expensive and unsuccessful plugins to be deactivated.
>
> But giving more detailed advice here is difficult without seeing the
> instance. You can send me a larger, problematic instance of your
> model as a personal message, and I will try to find the time to look
> at it.
>
> Best,
> Ambros
>
>
>
>
>
> Am 14.06.2018 um 23:21 schrieb Marcus Garvie:
>
> Hi everyone,
>
> this is my first post, so please understand that my knowledge of
> SCIP is low!
>
> I have been solving some large binary linear programming
> problems with no objective function. I solve .LP files using a
> terminal to issue the commands on my Mac. The .LP files are
> automatically generated, with e.g. the attached format.
>
> The only commands I issue are
>
> SCIP> read test.lp
> SCIP> optimize test.lp
>
> The problem is that for very large problems (e.g. 28456
> variables) the solver seems to be running forever (> 2 days)! I
> tried the same problem in Gurobi and it gave me the correct
> solution in 15 minutes. (I’m wanting to use SCIP because it has
> some easy options for giving me all feasible solutions, while
> Gurobi does not).
>
> I also tried the problem in CPLEX, but it has some limitations
> on the length of the variable names (Error 1464) so I’m a
> little stuck if I want multiple solutions for large problems.
>
> Any advice would be appreciated.
>
> Marcus.
>
> PS I only know how to problems in the .LP format!
>
>
>
> ________________________________
> Marcus R Garvie
> Associate Professor
> Rm 552 MacNaughton Bldg
> Dept. of Math & Stats
> University of Guelph
> Guelph, ON Canada N1G 2W1
> Tel. 519-824-4120 ext 53409
> Email. _mgarvie at uoguelph.ca_ <mailto:mgarvie at uoguelph.ca
> <mailto:mgarvie at uoguelph.ca>>
>
>
>
>
>
> _______________________________________________
> Scip mailing list
> Scip at zib.de <mailto:Scip at zib.de>
> https://listserv.zib.de/mailman/listinfo/scip
> <https://listserv.zib.de/mailman/listinfo/scip>
>
>
> --
> Ambros Gleixner, Research Group Mathematical Optimization Methods at
> Zuse Institute Berlin, http://www.zib.de/gleixner
> _______________________________________________
> Scip mailing list
> Scip at zib.de <mailto:Scip at zib.de>
> https://listserv.zib.de/mailman/listinfo/scip
> <https://listserv.zib.de/mailman/listinfo/scip>
>
>
>
>
> _______________________________________________
> Scip mailing list
> Scip at zib.de
> https://listserv.zib.de/mailman/listinfo/scip
>
--
Ambros Gleixner, Research Group Mathematical Optimization Methods at
Zuse Institute Berlin, http://www.zib.de/gleixner
More information about the Scip
mailing list